Joint Electricity Price and Load Forecasting for Data Center Battery Dispatch Using Polynomial Diffusion Models

summary

The gist

The paper proposes a real-time Battery Energy Storage System (BESS) dispatch framework that uses polynomial diffusion models to forecast both electricity prices and data center loads to minimize…

In short

This framework uses Polynomial Diffusion Models (PDMs) to create real-time forecasts for electricity prices and data center loads simultaneously. These forecasts feed into a dynamic programming system that optimizes battery charging and discharging to minimize procurement costs. The method significantly outperforms rule-based strategies, proving effective for managing energy costs in data centers.

Key concepts

Polynomial Diffusion Models (PDMs)
PDMs are used to model the random behavior of electricity prices and DC loads. They allow researchers to calculate expected values of polynomial functions efficiently using matrix operations, which is key for fast, conditional forecasting without needing complex simulations.
Dynamic Programming (DP) Framework
This optimization method uses forecasts to decide the best actions for the battery—when to charge or discharge. It finds the sequence of decisions that results in the lowest cumulative cost over time, respecting physical constraints like battery limits and power availability.
Conditional Interval-Averaged Forecasts
Instead of predicting a single future value, this method predicts an average value over a specific time interval. This is done by combining fixed seasonal patterns with stochastic residuals derived from the PDM to provide a more robust prediction for real-time dispatch.
Stochastic Residuals and Latent Factors
The random part of the forecast is modeled using latent factors that follow Ornstein–Uhlenbeck dynamics. These factors allow for efficient sampling of future uncertainty, which is then mapped into non-Gaussian processes to capture complex, asymmetric fluctuations in prices and loads.

Terminology used across episodes

This episode discusses

The paper

Joint Electricity Price and Load Forecasting for Data Center Battery Dispatch Using Polynomial Diffusion Models · Read on arXiv

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Joint Electricity Price and Load Forecasting for Data Center Battery Dispatch Using Polynomial Diffusion Models".

Dev: Here's the gist of the paper. The paper proposes a real-time Battery Energy Storage System (BESS) dispatch framework that uses polynomial diffusion models to forecast both electricity prices and data center loads to minimize electricity procurement costs Modeling Stochastic Dynamics Polynomial Diffusion Models (PDMs) are employed for forecasting both electricity prices and DC loads, capturing their respective stochastic dynamics through a unified framework The PDM property allows conditional expectations of polynomial functions to be computed via finite-dimensional matrix operations, which is crucial for efficient forecasting This approach extends PDMs from electricity price modeling to DC load modeling, suggesting that low-order polynomial transformations may adequately represent load variations Forecasting Methodology The conditional interval-averaged forecasts for electricity prices and loads are modeled as the sum of a deterministic seasonal component and a stochastic residual derived from PDMs Seasonal components are represented by truncated Fourier series containing daily and weekly harmonics, which are estimated offline The stochastic residuals are then modeled using mean-reverting latent factors governed by Ornstein–Uhlenbeck (OU) dynamics These latent factors admit an exact Gaussian transition, which facilitates efficient prediction and sampling without numerical discretization Finally, polynomial observation mappings transform the Gaussian latent factors into non-Gaussian processes to capture asymmetric fluctuations and extreme events Dynamic Dispatch Optimization The forecast-based dynamic programming (DP) framework optimizes BESS charging and discharging using the conditional price and load forecasts The Bellman recursion is used to find the minimum forecast-based cumulative cost from a given state, considering battery state transitions and degradation costs Constraints are imposed on power limits, energy ranges, and prohibit simultaneous charging and discharging operations The DP is solved every 15 minutes using the latest forecasts, implementing only the first optimal control action Performance Evaluation The proposed DP strategy achieves an aggregate cost savings ratio of 2.63 relative to rule-based (RB) dispatch across various BESS configurations Simulation results show that the DP strategy yields mean daily cost savings of C5,029/day compared to C1,959/day for the RB strategy The DP-to-RB savings ratio generally increases before leveling off as storage duration increases, indicating that forecast-based dispatch better exploits longer-duration storage flexibility This framework provides a computationally tractable approach to managing electricity costs amid growing demand The paper successfully developed a PDM-based forecasting and DP framework for cost-effective BESS dispatch in data centers The framework enables efficient conditional forecasting of both electricity prices and power demand Future work plans to incorporate higher-order predictive moments to enable risk-aware BESS operation REFERENCES arXiv:2611.08470v1

eess.SY: 5 Nov 2026, Joint Electricity Price and Load Forecasting for Data Center Battery Dispatch Using Polynomial Diffusion Models United Nations Economic Commission for Europe, “Aligning data centres and large load growth with grid resilience,” United Nations Economic Commission for Europe, Task Force on Digitalization in Energy, Geneva, Tech.

Rosa: Rep.

Dev: ECE/ENERGY/GE.6/2026/5–ECE/ENERGY/GE.5/2026 3,

1: A Wierman, Z Liu, I Liu, and H Mohsenian-Rad, “Opportunities and challenges for data center demand response,” in Proceedings of the International Green Computing Conference (IGCC).

Rosa: Dallas, TX, USA: IEEE, 2014, pp.

Title and authors: Rosa: So we’re looking at this paper, "Joint Electricity Price and Load Forecasting for Data Center Battery Dispatch Using Polynomial Diffusion Models." It basically proposes using these polynomial diffusion models to predict both the electricity prices and the power loads coming from data centers.

Dev: Right. And the big idea is that by using those polynomial diffusion models, they can calculate conditional expectations in a very efficient way through finite dimensional matrix operations, which is key for making these real-time forecasts actually usable.

Taro: It seems like they are extending this concept from just predicting electricity prices to also modeling the power demand coming from data centers, which is a big move because those two things really influence each other when you're dealing with large loads and market volatility.

Rosa: Exactly. The paper suggests that this unified framework allows them to forecast both price and load jointly, which helps in creating a more accurate picture for the battery system than just looking at one or the other separately.

Dev: And they are using these forecasts inside a dynamic programming framework to optimize when the battery should charge and when it should discharge, aiming to minimize electricity procurement costs.

Taro: That optimization part sounds like where the real action is, figuring out the best control actions for charging and discharging given what those forecasts tell them about prices and loads.

Rosa: And they're showing that this entire approach results in a cost savings ratio of two point six three compared to a rule based dispatch strategy across different battery setups, which gives you a concrete number to compare against existing methods <ref:2611.08470#pg1>.

Dev: That two point six three times savings is what they found when comparing it to the rule based strategy, and they also showed that for longer storage durations, those savings ratio numbers actually tend to increase before they level off <ref:2611.08470#pg1>.

Taro: So it means that having a forecast-based approach allows you to really take advantage of storing energy for longer periods, which is important for handling sustained demand shifts in data centers.

Rosa: And the paper details how they build these forecasts by taking a deterministic seasonal component and adding a stochastic residual modeled with Ornstein Uhlenbeck dynamics, which then gets transformed into non gaussian processes using polynomial observation mappings.

Dev: That transformation step is interesting because it lets them capture those asymmetric fluctuations and extreme events that you just can't get from simpler models, while the OU dynamics part allows for an exact Gaussian transition for efficient sampling without needing to discretize everything numerically.

Taro: So they’re essentially using these advanced mathematical tools to model how the price and load are fluctuating around their average patterns, which is a necessary first step before you can even start optimizing the battery usage.

Title and authors: Rosa: Before we get into how they make these forecasts online, I want to talk about the improvements they suggest for this framework. They focus on making those forecasts more practical for real-time use and adapting them quickly as new data comes in.

Dev: The paper suggests using an unscented Kalman filter, or UKF, to estimate those unobserved latent states—the things governing the stochastic dynamics—and updating them every fifteen minutes using new observations.

Taro: That sounds like a way to handle the uncertainty of those latent factors without having to run a full simulation every time data arrives; it’s an online state update mechanism.

Rosa: They also highlight that they can generate a deseasonalized conditional expectation for either the electricity price or the load by evaluating the polynomial basis at the filtered state estimate and doing some matrix vector operations, which avoids using stochastic simulation for online forecasting.

Dev: So instead of running a complex simulation every time we get a new data point, you just use that filtered state to calculate an expectation directly via those polynomial representations, which should help keep up with the real-time loop rate.

Taro: That sounds like it tackles the latency issue by making the calculation computationally tractable for online updates, which is crucial when you need decisions made every fifteen minutes.

Rosa: Moving into the results section and looking at how they tested this, they validate their polynomial diffusion model by calibrating the price PDM using historical German intraday electricity prices and using publicly available Google data traces for the load PDM.

Dev: They compare their simulated trajectories against actual out of sample observations over a five day period, and they found that both models capture the main temporal variations in both price and load well enough.

Taro: But they also noted something interesting there; when looking at the standard deviations, the paper says the polynomial diffusion models actually overestimate variability, especially for electricity prices, showing six seven point nine versus five three five for those prices.

Rosa: That’s a bit of a caveat then; it means their model is predicting more ups and downs than what actually happens in the real world for the price fluctuations.

Dev: It's important to note that this overestimation of variability is particularly pronounced with electricity prices, which might affect how aggressively the dynamic programming framework reacts to those predicted swings.

Taro: So, while the models capture the main trends, they aren't perfectly modeling every little spike or dip in price data when compared against reality.

Rosa: The paper then lays out how they handle this uncertainty in their DP framework by generating distributions of realized cost savings based on those PDM trajectories using a Monte Carlo simulation approach.

Title and authors: Dev: That Monte Carlo part is what lets them estimate the distribution of the actual cost savings, which gives you a fuller picture of the risk involved in making the dispatch decisions.

Taro: So they aren't just giving you one expected saving number, but a range or a probability distribution of what that saving might actually be in practice.

Rosa: And finally, we need to look at how they set up the actual battery control problem itself, detailing the physical constraints on charging and discharging.

Dev: They define the battery state of charge as being within a minimum and maximum range, say from bmin to bmax, and they describe how that state changes based on the charging power uc and discharging power ud over a time interval delta t.

Taro: The control set they are working with includes limits on the charging power uc, which is capped at u max c, and also limits on the discharging power ud, which is capped by min u max d L hat k.

Rosa: And there's that important constraint that you can't charge and discharge at the same time, meaning uc and ud cannot both be non-zero simultaneously.

Dev: That constraint is standard for battery systems, but they also include a cost term for battery degradation, which they approximate as a linear throughput cost c wear of uc plus ud multiplied by lambda times delta t.

Taro: So when you run the DP optimization every fifteen minutes, it’s looking at the current state of charge and deciding on those charging and discharging powers to minimize that total cost, including both electricity procurement costs and battery degradation costs.

Rosa: Overall, this paper with its title "Joint Electricity Price and Load Forecasting for Data Center Battery Dispatch Using Polynomial Diffusion Models" gives operators a way to integrate two key time-series forecasts into a single optimization problem.

Dev: It’s a computationally tractable framework because of those polynomial diffusion model properties, and they show that it can deliver cost savings of two point six three times the rule based strategy across various BESS configurations, which is pretty substantial <ref:2611.08470#pg1>.

Taro: It suggests that for data center operators, having this kind of forecast-based dispatch method is a viable way to manage electricity costs amid growing demand without needing to run incredibly complex and slow simulations constantly.

Rosa: So to wrap up, the authors developed a PDM-based forecasting and dynamic programming framework that lets you efficiently predict both electricity prices and DC loads for cost-effective BESS operation in data centers.

Dev: The paper’s real strength seems to be how they combine the polynomial diffusion model’s closed form conditional expectations with the dynamic programming optimization structure.

Taro: And the main implication is that it provides a method for cost-effective BESS dispatch by giving operators a way to handle electricity procurement uncertainty in a data center setting.

The paper's summary: Rosa: So, to recap what we just looked at about these papers, this new work is about building a real-time system that looks at both electricity prices and the power loads from data centers together using polynomial diffusion models to minimize costs.

Dev: Right. Basically, they take those two things—price and load—and model their randomness using a unified framework with these polynomial diffusion models so they can make these forecasts efficiently.

Rosa: And then they plug those forecasts into a dynamic programming system that decides when the battery should charge and when it should discharge to keep the electricity cost low. It’s all about making that decision fast and smart based on what those models predict.

Dev: That's exactly right, but what's really interesting is how they handle the uncertainty in those forecasts using an unscented Kalman filter to update their internal states every fifteen minutes with new data.

Rosa: That means even when the world changes unexpectedly, the system can keep updating its best guess without having to re-run a massive simulation every time something new comes in. It’s about being responsive in real time.

Dev: And they show that this method actually gets you a cost savings ratio of two point six three compared to a standard rule based dispatch strategy across different battery setups, which is a pretty solid metric for comparing it to existing methods.

Rosa: That number is significant because it shows the benefit of using these sophisticated models over simpler, more basic decision-making rules when you're dealing with data center energy management.

Dev: And they also point out that the savings ratio tends to get better as you let the battery store energy for longer periods, which suggests that this framework really helps exploit those longer storage opportunities.

Rosa: It means for someone just listening, this paper shows a path toward managing electricity costs in big data centers without needing to run incredibly slow and complicated simulations constantly.

Dev: And while they have their caveats—like the polynomial diffusion models sometimes overestimate how much variability there is in the price—the core idea is that you can integrate these forecasts into a dynamic programming framework for cost-effective battery operation.

Rosa: So, we’ve seen how they build these models and what kind of cost savings you get, but now we need to look at the actual experiments to see how well this whole thing holds up when it's running in the real world.

The paper's improvements: Rosa: We’re looking at how they suggest making this whole forecasting and dispatch system more useful in practice, right? It’s not just about having a model; it’s about making sure that model is actually running reliably when the world throws weird stuff at it.

Dev: Exactly. They focus on using an unscented Kalman filter to estimate those hidden states—those underlying factors driving the randomness—and they say you should update those estimates every fifteen minutes with any new observations you get.

Rosa: That’s key because it keeps the system current; it means that when a new price spike or a sudden load shift happens, the forecast gets corrected quickly instead of staying stuck on an old prediction.

Dev: And they also propose using those filtered state estimates to generate the deseasonalized conditional expectation for either the electricity price or the load through polynomial basis evaluation and matrix vector operations, which avoids needing a full stochastic simulation every single time.

Rosa: That’s smart because if you don't need to run a huge simulation every few minutes just to get an updated forecast, then you can actually deploy this on real hardware in data centers.

Dev: The implication there is computational tractability; it makes the whole process much more efficient for a real-time loop, which is what we care about when we talk about latency and failure modes.

Rosa: And they also mention that they need to use Monte Carlo simulation to show you the distribution of actual cost savings, so you get a range of outcomes instead of just one single number for the saving.

Dev: That gives a much better safety case because you’re not relying on a single expected value; you’re seeing the probability distribution of what actually happens when it goes live.

Rosa: So, to wrap up on the improvements, they are pushing for an online state update mechanism and using those updated estimates for fast expectation calculations to keep the dispatch loop tight.

Dev: It sounds like they are trying to bridge that gap between a theoretically perfect model and a system that can actually survive in an operational environment where data keeps flowing.

Rosa: And once you have these refined forecasts, what does this mean for the future? It sets up the next step, which is incorporating higher-order predictive moments into the models so they can handle risk more explicitly.

Conclusion: Rosa: So, to wrap things up on this paper, "Joint Electricity Price and Load Forecasting for Data Center Battery Dispatch Using Polynomial Diffusion Models," they’ve shown a way to link electricity price predictions with data center load predictions using these polynomial diffusion models.

Dev: Right. The main implication is that for data center operators, this gives them a computationally tractable method to manage electricity costs by feeding both forecasts into a dynamic programming framework for battery charging and discharging.

Rosa: It means they can move away from simple rule-based decisions toward something much more flexible that actually exploits the cost savings from storing energy over longer periods.

Dev: And the number is still pretty solid, showing that this approach yields two point six three times the cost savings of a rule based dispatch strategy across different battery configurations.

Rosa: That’s a concrete result for someone looking at how much capital they might invest in storage systems for data centers.

Dev: It’s also important to remember the caveats they mentioned; those polynomial diffusion models sometimes overestimate how much variability is there, particularly with electricity prices.

Taro: I agree with that, and it makes sense because if you overestimate volatility, your system might charge or discharge too aggressively when the price swings are less extreme than predicted.

Rosa: And they also pointed out that their method stops working in certain situations where the world misbehaves; they haven't fully tested what happens when things get really unusual outside of their lab trials.

Dev: Exactly, so while it’s a strong framework for predictable operations, we still need to figure out the robust safety case for what happens when those real-world conditions push the battery constraints to their limits.

Taro: That’s right; graceful recovery is what we need to test next, seeing how this framework handles those unexpected failures in a live environment.

Rosa: Well, this paper lays out a really clever way to combine forecasting and optimization for cost management in data centers using polynomial diffusion models.

Dev: It’s a solid piece of engineering that shows how stochastic dynamics can be integrated into real-time control loops effectively.

Taro: I'm curious to see how they expand this work to include those higher-order predictive moments so the system can actually anticipate more complex risk scenarios.

More episodes

← Home